Make these algorithms work for llms
If we operate these “outside” algorithms for LLMS, things are going badly. Therefore, we have reached improvements on algorithms that determine the main problems of operating “outside the box”.
For ELS, we had to go from DP guarantees at example to DP guarantees at the user level. We found it Previous work He was adding more noise orders than it was already necessary. We were able to prove that we can add a much lower noise, making the model much better while maintaining the same privacy guarantees.
For both ELS and ULS, we had to know how to improve the contribution binding. The “virtual” option is to choose a binding contribution for every user who is already satisfied with; That is, we do not do any prior treatment. However, some users may contribute a large amount of data, and we will need to add large quantities of noise to provide privacy to these users. It reduces the determination of a smaller contribution than the amount of noise that we need to add, but the cost should ignore a lot of data. Since LLM training is expensive, we cannot try training a set of models with different borders to contribute and choose the best – we need an effective strategy to choose a binding of contribution before We start training.
After widely prolonged experiment, we found that the associated contribution mode is that the average number of examples kept by each user is an effective strategy. For ULS, we offer a prediction of the overall noise as a function to link the contribution, and we found that choosing a contribution associated with reducing this prediction is an effective strategy.
